Giuseppe Barbalinardo

Great question. I started my PhD in physics 15 years ago, and I was like, you know, I'll finish grad school and I'll head back to Italy right after. And then one year after, I was like, yeah, you know what, I'll stay another year. And then two years after, I'll stay another year. And now it's 15 and counting. So I guess that's how I ended up, first in San Diego and then in the Bay Area.

It was a fun journey, especially with what I was studying in terms of material science and material simulations. I was focusing on semiconductors, and eventually you start getting smart, and you start using machine learning, you start using AI, you start making things work well with the resources you have, increasing the scale and so on. And that was the first effort that I did in terms of machine learning. Then I did a few years in the Bay Area as a software engineer, so this HPC experience helped me transition into software. And the last six years now, I've spent them at Tonal, where I still work in the AI field, but studying what I like to think is the most complex and interesting and beautiful physical system, that is, the human body.

Well, that was actually one of the reasons why I joined the company. There's this multi-scale understanding. Initially you start with sensors, high-frequency sensors. There's your measurement of the position of your arms, of the force applied to your arms. You have this measurement 50 times every second. You have a camera that understands your movement with 500 points every second, and so on. And then you need to convert that into a user journey, into an experience that lasts months, into a human outcome, a long-term journey. You want to recover from an injury, you want to become stronger, you want to improve your mobility. All these long-term goals need to be encoded, and we need to start from the very elementary sensors and build this longer-term trajectory in a way that becomes meaningful and brings real outcomes. And this is somehow similar to what you do with atoms and molecules. You start with a few properties of silicon, of carbon and so on, and then you need to build the statistical properties, the thermal conductivity, the long-term predictions of fluctuation currents and so on.

Tonal is the most intelligent home gym, and we have a lot of sensors to make that happen. It's a device that you can install in your home, or it can be installed in a physical therapy studio, in a hospital, and help you recover. At the same time, it has a lot of intelligence. It has a lot of sensors that are not something you would have if you go to the gym. For example, if you are doing a workout at 4 a.m. and you're doing a bench press, you have a hundred pounds that you're trying to lift, and you don't have two friends helping you. Well, if you're doing it with the machine, there's something called Auto Spotter that detects vibrations, detects the speed, and a lot of signals that can be collected by these sensors, and then we apply machine learning on top of that, which reduces the weight, and you can safely finish your workout if you're struggling. At the same time, we can detect very precise changes, for example, in the symmetry of your shoulders, or very precise sub-inch changes in your range of motion. These are really signals that are difficult to understand if you're in a clinic and you're doing the measurements with a rubber band or with a goniometer. The reality is that the world is shifting towards what I like to think of as the operating system of your body, where you have a lot of sensors and a lot of metrics, biomarkers that you can track. And the part that Tonal can offer that is not, I think, as obvious, is the actionability part. It's not just measuring, like your ring or your watch. It actually is, okay, now go do this exercise, this session, this stretching, this mobility session, this hypertrophy workout, and so on. So really, it's closing the loop into the human outcome part.

Well, a lot. You can connect all sorts of devices, wearables. They all enter into the ecosystem that we're building. And there are some unique signals that are specific, provided by the machine. Tonal is not using regular weights. It's using electromagnetic resistance, and therefore we can measure a lot of things, for example, force, position, velocity, power, heart rate, all these signals during your activity, during your session, plus all the history that these sensors give you. And of course, I forgot to mention the camera, which is then processed into key points, and these landmarks are then processed into biomechanics, and these biomechanics are augmented with the force measurements. So this capability of understanding movement, of understanding your unique biomechanics, together with these other signals, can build this unique fingerprint of your performance, unique fingerprint of your mobility, unique fingerprint of your endurance. And together with some of the structured data that we collect, for example, what are your goals, what's your age, we have a lot of longitudinal data. We can study, say, I'm a 40-year-old male living in San Francisco. We have a lot of segments of data, cross-sectional, so we can study a whole segment of people similar to me and understand, with my specific fingerprint, how I can optimize this performance.

Exactly, yes. It's a very interesting and amazing period, where we have a ton of data. Of course, from these millisecond signals multiplied by hundreds of thousands of devices, we have a ton of data. Therefore we can use this data, and use your unstructured input, because you can use TIA, as we call it, your virtual coach, to provide other data. For example, oh, I'm tired today because I went for this activity, or I didn't sleep well. Or add information that doesn't come from wearables, that is specific to you. This unstructured data, plus structured data, plus sensor data, and of course compared to all the rest of the population and people similar to you, really creates a unique recommendation for your next five minutes, for your next day, for your next month.

More technologies are developing that allow us to measure things that before were not possible. Think about glucose monitors. That's something where now you can have tracking from after your breakfast to when you go to bed, and you have this signal now, and it's super important because it tells you when you have more energy, when you have less energy, and so on. Now there are hormone trackers coming that are able to measure, some of them more reliably, some of them less reliably, the level of cortisol during the day, or levels of other factors that before you would have no idea how to measure. And I think this is really opening up opportunities, not only for Tonal, but for us as consumers. The whole medical healthcare system is going through this revolution where you now have a way to, first of all, optimize your life in this moment, but also start to predict how you will feel in a minute, how you will feel in one hour. For us, for example, predicting injuries is a very interesting area. How much a signal of asymmetry, for example, could be an indicator of an injury coming. And up to a certain point, if you have enough input data and enough labels, and unfortunately labels are very expensive, the architecture, yes, it is important, and yes, you need to study and develop the right one, but it's also something where eventually you get to figure out the right architecture. So really, the value, and I think this is true in AI in general, even when we think about LLMs, the value is really the input data that you're putting into the system.

And therefore, I think it's up to us, first of all, to create the right setup, the right infrastructure to do so, and up to us to make it safe. And maybe that's a part where we should double-click. One of the challenges that we needed to solve when we built our intelligent coach was around injuries. We do not want to provide medical advice. We want to make sure that there are boundaries on the knowledge that the chatbot provides to you. In general, I feel like it's not always the case. For us, it was a big effort. I don't know if all companies do that, especially the smaller ones, but I think it's something we should be very careful about when we try a new product. Like, okay, what kind of data are these models trained on? Are they able to provide the recommendation for the specific information they're giving me? Is that something that they should be doing, or should they not be doing it and they're doing it anyway?

Yeah, the diversity of data. Again, I don't know how close we are to that point, but I think at some point we'll really start to think about not the Anthropic LLM, Claude, or the OpenAI LLM. We'll start thinking about what's the underlying data for this specific purpose, this LLM designed for this specific purpose. And for us specifically, in terms of humans and what Tonal does, we have a pretty diverse population, and there are techniques for making sure that you don't introduce bias, basically. I remember at the beginning, when you were asking ChatGPT to generate a watch, it would generate this 10:10 time on the watch every time, because historically all the images that were produced of watches were always showing this time, because it looked beautiful. You had your Rolex in the middle and then this framing through the time. That was clearly a bias. A very visible bias. Now, how do you correct for other biases that are less visible? That is really a challenge. And of course, when you do your training, when you do your new model, it's in the hands of the developers. There are techniques for doing that, but it needs to be something that you design for. It's not something that comes automatically, and it should not be taken for granted. We have a strong policy about data sharing and data collaborations, and in fact, we don't sell data. In general, I think that's kind of the trend for a lot of companies. In the health space it's even more important, but I would say for a lot of companies now... For example, in the past they had open APIs and easier access to data. Now everybody's becoming a little more attached to their data. And of course that's also because your value as a company, yes, of course, is the hardware. Software is a little more commoditized now, and data is the value of your company. So the moment that the data leaves the door, well, you lost a lot of the value of the company. Plus, of course, as you were touching on, the ethical part behind that is perhaps more important.

There are two departments focused on this: software, and my department, which is data science and AI. My team is around 20 people total. I say a small but mighty team. We deal with all sorts of data engineering, sensors and intelligence, analytics, and then research. So we have four main areas of focus. For example, one of the most recent works there was this AI coach that we've been developing, which covered pretty much all four aspects, starting from the research part. How do you design that? How do you create the right harness, the right skills, the right access to the data? How do you need to refresh data? When I think about the design of a new feature, at least modern features, I always imagine this iceberg picture, where there's the tip of the iceberg, which is the part of the feature that you see, and then there's a whole huge hidden part: how do you manage the tracing, the safety of the data, the compliance that you need to take care of, cost, lineage, the infrastructure. And this was a lot of the work that my team had to do together with the software department, because a lot of the underlying infrastructure was built by the software department. So really, there's a lot that goes beyond what you see when you open the app and you're asking a question to TIA, which is the name of the feature, and you get the answer. I think very few people know how many evals needed to go behind that answer, how many traces needed to go behind that answer, or even just cost tracking and optimizations.

That's a great question. I feel like things are very dynamic. They are still evolving every day, and we haven't found yet the best structure for collaborating. I heard this term, "meat proxy," in the past. There was a funny one that I think was from the OpenClaw guy, where he was saying that at this point, sometimes the conversation between me and you is actually between my agent and your agent. However, it goes through me and you, in a way that I ask my agent and then I reply to you, and then you ask your agent and you reply to me, and so on. So let's say this is one of the open problems. When we open a pull request, does your AI write the description of your pull request when you add a new piece of code that is used for developing a new feature? Do you write the description on the pull request, or do you include the whole conversation that your agent had, or do you include the link so that you can continue that conversation? These are great open questions. I don't have the answer today. I know that it's something that we're experimenting with. I don't know if other teams have.

Yeah, it's a great question. I think the dynamics are changing, in the sense that we are embedding a lot of the human process that we were following as humans into the AI now. For example, you have sprint planning, or whatever you're using, agile or whatever methodology you're using for managing your team. Now you can delegate that to AI. Now, it's still the human. The human's still the one you need to blame, the one with the responsibility. So even if I use my AI to review 30 pages and create bullet points, if I was in charge of reviewing that document, then as of now, at least, I don't know how the future will be shaped there, but as of now, I'm still the person responsible for that. So if I didn't read it, it's on me. If I put my signature down there and I approved it, it's on me. And I think this is one area that was like the North Star for my team. Okay, the goal is to improve the process. The goal is to improve productivity. And the goal is to be smart about leveraging this revolution that is happening. At the same time, let's not forget, number one, the responsibility, as of now at least... I don't know, at some point there will be insurance for AI, or there will be a new world of delegation of responsibility. But as of now, the responsibility is still on us humans. Number two, one of the advantages of leading the data science and AI team is that they all come from a world of machine learning. Machine learning is non-deterministic in a sense, but it's also deterministic, because when you write a model, you write the evals for that model. You write the training, and you split the data set into training and testing, write a validation there. Now you write the evals. You design the model in a way that is measurable somehow. You design it as a state machine, and you design it in a way that can be traced. And the same can be done when you create a skill for your agents. You create a Jira skill, or you create a data layer for letting your org access the data as a self-serve tool. And that means that as you design the infrastructure for that, you need to design the counterpart, the validation, and make sure that you know exactly that today 95% of your queries are correct, and there's this 5% that is not correct. Maybe because your semantics or your ontology doesn't cover all these details. What's the definition of an active user? These definitions can be not as precise. And I think really it's up to us to approach this problem, this opportunity, in a scientific way, where you can make everything measurable, including how you optimize your tools, including how you measure things, including how you design a new process.

Yeah, exactly that.

So far we've been growing. We currently have an open role in the research area, and we are definitely interested in building more technology, building more features, but also expanding the way we leverage data, expanding the way we improve outcomes. There is a little bit of this fear of, okay, AI will replace us all. Well, of course there's never a black or white answer. But at the same time, if we are able to evolve together with the new tools that are provided to us, I see it as, if anything, having superpowers. Maybe tomorrow my job will be just chatting with the agent without even having a screen, or who knows how the future will be. But at the same time, I think as long as you keep yourself up to date and keep using new tools, especially in this critical moment where things are changing so fast, I feel like that's the requirement that allows you to not become old, in a sense, and for your job not to evaporate.

Thank you.